Skip to content

[Bugfix][ROCm] Fix startup OOM in AITER MLA FP8 prefill workspace sizing #57923

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
dllehr-amd merged 2 commits into from
Sep 23, 2026
Merged

[Bugfix][ROCm] Fix startup OOM in AITER MLA FP8 prefill workspace sizing #57923

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5598b5b
fix: size upper bound max num partial tiles manually
simondanielsson Sep 21, 2026
db2a03f
Merge branch 'main' into fix/aiter-asm-prefill-sizing
simondanielsson Sep 23, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion vllm/v1/attention/backends/mla/rocm_aiter_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -772,9 +772,18 @@ def _init_fp8_prefill_ps_buffers(
device=device,
)

from vllm.platforms import current_platform
from vllm.v1.worker.workspace import current_workspace_manager

max_num_partial_tiles = reduce_partial_map_size
# AITER metadata sizing assumes all requests can carry max_prefill_qlen tokens,
# which is a loose worst case in the number of QO tiles. Rather, the sum of
# qlens is bounded by max_num_batched_tokens. Manually compute the number of
# qo tiles to avoid OOM at startup.
# TODO: AITER should give us this budget constrained value
qo_tile_cnt = (
cdiv(max_num_batched_tokens, _FP8_PREFILL_TILE_Q) + max_num_reqs - 1
)
max_num_partial_tiles = qo_tile_cnt + current_platform.num_compute_units()
reservations: list[tuple[tuple[int, ...], torch.dtype]] = [
(
(max_num_partial_tiles * _FP8_PREFILL_TILE_Q, num_head_k, v_head_dim),
Expand Down
Loading